Logarithmic Neural Network Accelerator Addition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network computations, particularly in convolution layers, face energy inefficiencies due to the complexity of performing accurate addition operations on logarithmic format values, which often require conversion to integer format and back, leading to increased energy consumption and time.
Innovation Solution
A method and system that decompose logarithmic format values into quotient and remainder components, allowing for efficient addition by sorting and summing these components, reducing the need for frequent conversions and thereby lowering energy consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If logarithmic format values are added by converting to integer format and back, then addition can be performed, but energy consumption increases and computation time increases
Solution Approach 1:
The patent segments logarithmic format values into quotient components and remainder components. This segmentation allows the addition operation to be performed directly on the components without converting to integer format, thereby reducing energy consumption while maintaining the ability to perform addition efficiently.
2Measurement precision
If more bits are used to represent each value, then calculation accuracy improves, but energy consumption increases and computation time increases
Solution Approach 1:
By segmenting logarithmic values into quotient and remainder components, the patent enables accurate calculation using fewer bits. The quotient component captures the magnitude while the remainder component captures the fractional part, allowing precise representation without requiring additional bits, thus reducing energy consumption.
3Use of energy by moving object
If logarithmic format values are decomposed into quotient and remainder components and sorted, then addition becomes more efficient and energy consumption reduces, but device complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-sorting the quotient components based on remainder components before performing addition. This preliminary organization of data allows the addition operation to proceed efficiently without requiring complex real-time comparisons during the addition process itself, thereby reducing overall device complexity.
4Ease of manufacture
If conventional addition methods are used for logarithmic format values, then implementation is straightforward, but bit toggling increases and energy efficiency decreases
Solution Approach 1:
The patent segments logarithmic values into quotient and remainder components, which can be processed independently. This segmentation reduces bit toggling because the quotient components (which are sorted) and remainder components can be added with fewer bit changes, thereby improving energy efficiency while maintaining implementation feasibility.
Data Source
AI summary
Neural networks, in many cases, include convolution layers that are configured to perform many convolution operations that require multiplication and addition operations. Compared with performing multiplication on integer, fixed-point, or floating-point format values, performing multiplication on logarithmic format values is straightforward and energy efficient as the exponents are simply added. However, performing addition on logarithmic format values is more complex. Conventionally, addition is performed by converting the logarithmic format values to integers, computing the sum, and then converting the sum back into the logarithmic format. Instead, logarithmic format values may be added by decomposing the exponents into separate quotient and remainder components, sorting the quotient components based on the remainder components, summing the sorted quotient components to produce partial sums, and multiplying the partial sums by the remainder components to produce a sum. The sum may then be converted back into the logarithmic format.